Systems and Methods for Multilingual Text Generation Field

1. A system for concurrently generating parallel texts in at least two different languages, the system comprising: one or more processors; one or more computer-readable media storing computer-executable instructions; and wherein the instructions, when executed by the one or more processors, causing the one or more processors to implement a neural machine translator and a generative adversarial network; wherein the neural machine translator comprises: a shared latent space including a plurality of coded representations, each of the plurality of coded representations mapping to a set of parallel texts in the at least two different languages, wherein each set of parallel texts includes texts in each of the at least two different languages, wherein each text included in each set of parallel texts has the same meaning; and an autoencoder comprising: an encoder configured to receive sample text in a first language of the at least two different languages and output a coded representation from the plurality of coded representations of the shared latent space which maps to one of the sets of parallel texts that includes the sample text in the first language and text in a second language which has the same meaning as the sample text; and a decoder configured to receive the coded representation and output the text in the second language; and wherein the generative adversarial network comprises a generator and a discriminator, the generator configured to receive noise and generate a fake coded representation of text in the first language of the at least two different languages, conditioned on the noise, that mimics a coded representation of the plurality of coded representations included in the shared latent space, the discriminator configured to receive the fake coded representation from generator and the coded representation from the encoder and provide feedback to the generator for use in learning to generate the fake coded representations.

2. The system of claim 1 , wherein the autoencoder is pre-trained using a training dataset comprising texts in each of the at least two different languages and a loss function to learn the plurality of coded representations included in the shared latent space.

3. The system of claim 1 , wherein the discriminator is configured to: produce a confidence score representing how certain the discriminator is that the fake coded representation is the coded representation received from the encoder.

4. The system of claim 1 , wherein the generator and the discriminator each comprise a neural network.

5. The system of claim 4 , wherein the discriminator is configured to provide feedback to the generator for use in learning to generate fake coded representations by evaluating a discriminator loss and back-propagating a gradient penalty of the discriminator loss to optimize parameters of the discriminator and by evaluating a generator loss and back-propagating a gradient penalty of the generator loss to optimize parameters of the generator.

6. The system of claim 1 , wherein the encoder receives the sample text in a first language from a training dataset comprising sample texts in each of the at least two different languages that is used to train the autoencoder to learn the plurality of coded representations included in the shared latent space.

7. The system of claim 1 , wherein the encoder is further configured to receive sample text in the second language of the at least two different languages and output a coded representation from the plurality of coded representations of the shared latent space which maps to one of the sets of parallel texts that includes the sample text in the second language and text in the first language which has the same meaning as the sample text in the second language; and the decoder is further configured to receive the coded representation and output the text in the first language; and wherein generator is configured to receive noise and generate a second fake coded representation of text in the second language, conditioned on the noise, that mimics a coded representation of the plurality of coded representations included in the shared latent space, and the discriminator is further configured to receive the second fake coded representation from generator and the coded representation from the encoder and provide feedback to the generator for use in learning to generate fake coded representations.

8. The system of claim 7 , wherein the encoder and decoder each comprise a neural network.

9. The system of claim 7 , wherein the noise is sampled from a Gaussian distribution.

10. The system of claim 7 , wherein each text included in each set of parallel texts is a sequence of words or a sentence.